Why AI-generated text irritates Habr readers, and how to edit it by hand

A column has appeared on Habr AI about why text generated by neural networks irritates readers even when it contains no obvious errors. The author looks not at model bugs but at the editorial hallmarks of synthetic text that make readers lose attention and trust.

Why it's so maddening

The column's main thesis is simple: readers are annoyed not by the fact that AI was used, but by the text's empty confidence. A neural network can assemble a tidy piece with no obvious failures, an even structure and logical transitions, yet still fail to give the sense of real thought. It doesn't argue with the reader directly, but constantly acts as if it has already explained everything. As a result, even a paragraph that is fine in form sounds suspicious.

The author puts it bluntly:

a neural network too often writes text that looks like text but

doesn't feel like thought.

The problem shows up very quickly. After just two or three screens, such text starts to tire the reader: attention slides across the paragraphs, and only generic words stay in memory. Instead of an argument, the reader gets a smooth imitation of expertise. For a technical audience this is especially painful, because it quickly senses when a piece is built from safe wording but doesn't rest on experience, observation or a clear position from the author. The reader sees the form but not the support beneath it.

Where the text breaks down

The column lists typical signs by which an AI draft starts to irritate well before the end. It's not only about clichés. What works worst is long introductions, a universal structure and an attempt to sound weighty without enough specifics. If a paragraph can be cut in half without losing meaning, if the subheadings can easily be swapped around, and if conclusions look final where the topic is still contested, the reader notices almost instantly. The text looks assembled but doesn't move toward a new idea.

• A long run-up instead of substance
• A skeleton of universal subheadings
• Too even a rhythm of paragraphs
• Generic words without facts and scenarios
• A categorical tone without caveats

The author separately singles out a "translated" aftertaste. Many texts are formally written in Russian but sound like a poorly adapted English-language presentation. On Habr such things get spotted especially fast: the local audience reads a lot, has little patience for filler, and notices whether the author understands the subject or is simply assembling plausible text from familiar words. So the "generate, lightly polish and publish" approach works less and less well on such a platform.

How it's fixed

The author writes that he almost never treats an AI draft as a finished article. For him it is raw material from which a readable piece still has to be built. The first step is to remove all the routine warm-up and open the text with a fact, a conflict, an observation or a conclusion. Next, the logic is rebuilt: not "what it is, how it works, why it matters," but an order that actually carries the thought forward. Subheadings are also rewritten almost every time, because they should not divide the text but move it forward.

Tougher editing follows. Crutch phrases such as "an important aspect" and "opens up new possibilities" are scrubbed out, concreteness is restored to every general claim, the sterile rhythm of identical paragraphs is broken, and artificial categoricalness is toned down. If a model promises to save time, you need to show where exactly; if a service supposedly handles complex tasks better, you need context and a comparison. Cosmetic "humanizing" doesn't help here: a weak synthetic text is usually easier to rewrite from scratch than to polish on the surface.

What it means

The column neatly captures the limit of generative models' usefulness: they speed up the rough-draft stage, help you get moving and sketch a structure, but they don't replace an editor. For media outlets, corporate blogs and content teams, this is a direct signal: publication quality is determined not by whether AI can be detected, but by whether the text has density of thought, specifics, a natural tone and respect for the reader. Otherwise even a grammatically smooth piece quickly turns into irritating noise that looks convincing but leaves nothing behind after reading.